If I have information about a person e.g. their lifestyle, how big is their house, where they live, their income history, I can build a model to predict their future income. But what if my task is not to predict the future income, but predict yes/no questions related to income e.g.
"Will the income be between 50K and 100K during the next 5 years?",
"Will the person have lower income (e.g. 30K) before higher income (e.g. 150K) in the next 5 years?"
For example, if the guy lose his job, then he will become poor (has lower income e.g. 30K), and then he will get a new job with 150K income during the next 5 years. Then the answer to the question is YES.
All the numbers in the examples above must be adjustable. In training, all the answers can be generated because we know the true income of him in the training set. The reason I want to do this is to measure probability of YES. This probability is useful for making some decisions e.g. should you help the guy find the job or invest in him if the model predicts he will be poor before he will be rich?
If you want to train a model to predict this, you must feed all the parameters of the questions as input as well.
Here is example training process:
- generate random lower and upper bound e.g. 50K and 100K and also the random year e.g. 5 years
- generate the label by checking whether the guy's future income is between 50K and 100K during next 5 years. If the guy's future income is 75K then the label is YES
- train the model with all the features, lower bound, upper bound, and year as input query, to predict the label
It's like a data augmentation process where you will have a lot of data samples from few training samples you have.
Is there any general paper/artcle/techniques that does ML for yes/no questions like this?
Or do you have a suggestion about how to build such model?
My naive idea sounds like it should work, for each training sample, you can generate like 100 samples from one training sample. But if there's like someone researching this idea already it would be better.